Meta Wants You to Tell Facebook's Feed What to Show
Meta is rolling out explicit personalization for Facebook Feed recommendations, asking users directly what they want the ranking system to show them. Engadget
The control lives in a new "your algorithm" menu attached to individual Feed posts and Reels. Access starts at the item level. You can open the menu from a specific post or Reel and adjust preference signals that shape later recommendations.
The menu brings three inputs together in one place. You can add specific topics to an interests list. You can add specific topics to a "not interested in" list. There is also a free-text box that asks you to describe in your own words what you want to see in Feed.
Meta says the feature is possible because its ranking systems now use large language models, AI systems trained to understand everyday language. That detail shapes how the inputs work. Free-text statements, topic inclusions, and topic exclusions must be interpreted for meaning before they can act as ranking signals. Rankers built this way can accept that kind of unstructured, user-written input directly.
The company introduced a similar algorithm personalization feature on Instagram earlier. The Facebook rollout extends the same approach of stated interests, stated disinterests, and plain-language description to Feed and Reels recommendations.
In my view, the technical interest is the shift from inferred to declared intent. Feed ranking has long relied on behavioral traces such as watches, likes, time spent, and skips. Those signals offer volume but they are noisy. They mix exposure with endorsement and leave the system to guess why you lingered or scrolled past. It is like telling a cook what you like rather than having the kitchen guess from what you left on the plate. A declared interests list and a "not interested in" list give cleaner labels. They trade volume for precision.
Worth flagging is how plain-language input changes the tuning problem for practitioners. Prompt-style input lets users state combinations, exceptions, and degrees of interest that checkboxes handle poorly. "More trail running, less marathon training content" is easy to write and hard to fit into a fixed taxonomy, or preset list of categories. For a language-model ranker, that sentence can be used directly as context when scoring posts. The cost is ambiguity. People write vaguely, contradict themselves over time, and revise preferences without clearing old entries. Any working system must resolve conflicts between the free-text statement and the structured lists, and between new statements and accumulated history.
Another practical question for Feed builders is control granularity, or how broad a preference should be. Item-level entry makes sense for capture. Frustration or delight happens at the post. Whether that preference should apply to a topic, a format, a creator, or a passing mood affects precision and recall, the balance of relevant results shown and irrelevant results avoided, across the pool of candidates. Explicit controls cut guesswork, but they ask maintenance work from the user. Most people will not curate lists carefully. The systems that gain most will treat a short statement as a strong starting point, then keep refining with behavior rather than freezing on the literal wording.
The broader context here is user agency as a product surface, or controls people can see and edit. Giving people readable, writable handles on recommendation logic does not replace ranking work. It adds a valuable new input alongside embeddings, engagement predictors, and retrieval stages, the existing steps that match content to users and fetch candidates. When it works, the payoff is faster agreement on what a specific user means by relevant, with fewer wasted impressions. That is a practical gain for daily use, and a useful direction for recommendation design to keep exploring.
